syntaxsynth/Ultra-FineWeb-L3-zh-hant-translated

Ultra-FineWeb-L3 (Traditional Chinese Translation)

Traditional Chinese translation of the English openbmb/Ultra-FineWeb-L3 corpus, targeting Taiwan-standard Traditional Chinese (臺灣正體中文).

Background

Ultra-FineWeb-L3 is the L3-refined tier of the UltraData data management framework developed by OpenBMB. Starting from Ultra-FineWeb (the high-quality web corpus behind MiniCPM4), L3 refinement transforms raw web documents into two structured synthesis formats:

• —Q&A Pair Generation — each document is converted into an "original text + multiple question-answer pairs" record, covering conceptual understanding, causal analysis, comparison, inference, and synthesis question types.
• —Multi-style Rewriting — each document is rewritten into four distinct styles (encyclopedia, textbook, blog, abstract), reorganizing core knowledge for richer training signal diversity.

This dataset is a full translation of the English L3 output into Traditional Chinese, making it suitable for training or fine-tuning language models on Traditional Chinese text with high knowledge density and structural diversity.

Translation Methodology

The system prompt instructs the model to match the tone and register of the source, preserve all structural formatting, keep code/math/URLs untranslated, and output only the translated text with no commentary.

Subsets

multistyle — 2,495,221 examples

Multi-style rewrites of web documents. Each record contains one rewrite in one of four styles:

The subset also includes a small number of blog-style records synthesized from a Traditional Chinese local corpus (TMMLU+), globally shuffled with the translated records.

qa — 1,194,156 examples

Q&A pairs generated from web documents. Each record contains the original document prepended to multiple structured question-answer pairs. Q&A pairs cover five question types: conceptual understanding, causal analysis, comparison/contrast, inference/application, and synthesis/summary.

Schema

Both subsets share the same flat schema:

{
  "uid":  string,   # unique document identifier (inherited from source)
  "text": string    # translated / synthesized Traditional Chinese text
}

Usage

from datasets import load_dataset

# Multi-style rewriting subset
ds_multi = load_dataset("syntaxsynth/Ultra-FineWeb-L3-zh-hant-translated", "multistyle", split="train")

# Q&A pairs subset
ds_qa = load_dataset("syntaxsynth/Ultra-FineWeb-L3-zh-hant-translated", "qa", split="train")

Disclaimer

This dataset is provided as-is. Translations are machine-generated and have not been manually verified. Quality may vary across documents, particularly for highly technical, domain-specific, or ambiguous source texts. Users should apply their own quality filtering as appropriate for their use case.